NLP Lineage

6 July 2023 · Representation and retrieval

Lost in the middle

Lost in the Middle: How Language Models Use Long Contexts

N. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, P. Liang

Date
6 July 2023
What the date marks
arXiv submission
Kind
Research paper
Organisation
Stanford / Berkeley / Samaya AI
Access
Not applicable
Other dates
TACL 2023
Threads
In the guided tour
Chapter 10. Reading more than you remember
Learning phases
4. Embeddings and RAG

Notable: no later milestone in the data is documented as descending from it. How levels are decided

Read the primary source (arxiv.org)

Full summary in progress. Until then, this page shows the milestone's record from the research data.

What it changed

Before it

Longer context windows were assumed to mean more usable information.

The problem

Nobody had checked whether models actually use the middle of a long prompt.

What it introduced

Accuracy is high when the relevant passage is at the start or end and sags badly in the middle, in a U-shaped curve.

Why it matters

It explains why stuffing everything into context underperforms good retrieval, and why reranking still matters.

What modern models inherit

The knowledge that context position affects whether information is used.

What came next

Better long-context training and evaluation, and continued investment in retrieval.

Before and after

The milestones it descends from, and the milestones that descend from it. Work developed alongside it is not counted. Weak links are left out unless you include them.

Before: where it came from

Lost in the middle has 24 documented ancestors across 5 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 6 turning points among them.

Lost in the middle has 24 documented ancestors across 5 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

Lost in the middle has 41 recorded ancestors, including indirect ones, across 9 chapters, reaching back to Recurrent networks in 1990. Showing its direct parents and the 12 turning points among them.

Lost in the middle has 41 recorded ancestors, including indirect ones, across 9 chapters, reaching back to Recurrent networks in 1990. Showing all of them.

  1. 2023Lost in the middleThis milestone.
  2. 2020RAGLost in the middle was influenced by thisChapter 10. Reading more than you remember
  3. 2020Dense retrievalRAG built directly on thisChapter 10. Reading more than you remember
  4. 2019BARTRAG built directly on thisChapter 4. Pretrain once, use everywhere
  5. 2018BERT Turning pointDense retrieval built directly on thisChapter 4. Pretrain once, use everywhere
  6. 2018GPT-1BERT was influenced by thisChapter 4. Pretrain once, use everywhere
  7. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  8. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  9. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  10. 2017Transformer Turning pointBERT built directly on thisChapter 3. Attention, then only attention
  11. 2017FAISSRAG was influenced by thisChapter 10. Reading more than you remember
  12. 2016Layer normTransformer built directly on thisChapter 3. Attention, then only attention
  13. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  14. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  15. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  16. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  17. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  18. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  19. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  20. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  21. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  22. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  23. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  24. 1997LSTM Turning pointELMo built directly on thisChapter 1. Language as probability
  25. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability
  1. 2023Lost in the middleThis milestone.
  2. 2023100K contextLost in the middle was influenced by this (indirect influence)Chapter 10. Reading more than you remember
  3. 2023Claude100K context built directly on thisChapter 7. The assistant goes public
  4. 2022Constitutional AIClaude built directly on thisChapter 6. Teaching models what people want
  5. 2022HH-RLHFClaude was influenced by thisChapter 6. Teaching models what people want
  6. 2022Chain of thought Turning pointConstitutional AI was influenced by thisChapter 12. Thinking longer
  7. 2021HHH assistantHH-RLHF built directly on thisChapter 6. Teaching models what people want
  8. 2020RLHF for summarisation Turning pointHH-RLHF built directly on thisChapter 6. Teaching models what people want
  9. 2020GPT-3 Turning pointChain of thought was influenced by thisChapter 5. Scale becomes a science
  10. 2020RAGLost in the middle was influenced by thisChapter 10. Reading more than you remember
  11. 2020Dense retrievalRAG built directly on thisChapter 10. Reading more than you remember
  12. 2020Scaling laws Turning pointGPT-3 was influenced by thisChapter 5. Scale becomes a science
  13. 2019BARTRAG built directly on thisChapter 4. Pretrain once, use everywhere
  14. 2019RLHF for LMsRLHF for summarisation built directly on thisChapter 6. Teaching models what people want
  15. 2019Sparse Transformer Turning pointGPT-3 built directly on thisChapter 4. Pretrain once, use everywhere
  16. 2019GPT-2 Turning pointGPT-3 built directly on this
    RLHF for LMs built directly on this    Chapter 4. Pretrain once, use everywhere
  17. 2019WebTextGPT-2 was influenced by thisChapter 4. Pretrain once, use everywhere
  18. 2019Transformer-XLSparse Transformer was influenced by this (indirect influence)Chapter 4. Pretrain once, use everywhere
  19. 2018BERT Turning pointDense retrieval built directly on thisChapter 4. Pretrain once, use everywhere
  20. 2018GPT-1BERT was influenced by thisChapter 4. Pretrain once, use everywhere
  21. 2018ELMoBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  22. 2018ULMFiTBERT was influenced by thisChapter 4. Pretrain once, use everywhere
  23. 2017CoVeELMo was influenced by thisChapter 3. Attention, then only attention
  24. 2017RLHF originConstitutional AI was influenced by thisChapter 6. Teaching models what people want
  25. 2017Transformer Turning pointBERT built directly on thisChapter 3. Attention, then only attention
  26. 2017ConvS2STransformer was influenced by this (indirect influence)Chapter 3. Attention, then only attention
  27. 2017FAISSRAG was influenced by thisChapter 10. Reading more than you remember
  28. 2016Layer normTransformer built directly on thisChapter 3. Attention, then only attention
  29. 2015Residual connectionsTransformer was influenced by thisChapter 3. Attention, then only attention
  30. 2015LM pretraining Turning pointBERT was influenced by thisChapter 2. Words become vectors
  31. 2015BPE tokenizationGPT-2 built directly on thisChapter 2. Words become vectors
  32. 2015Luong attentionTransformer was influenced by thisChapter 3. Attention, then only attention
  33. 2014AdamTransformer was influenced by thisChapter 2. Words become vectors
  34. 2014GloVeELMo built directly on thisChapter 2. Words become vectors
  35. 2014Seq2seq Turning pointTransformer built directly on thisChapter 2. Words become vectors
  36. 2014Attention Turning pointTransformer built directly on thisChapter 3. Attention, then only attention
  37. 2014GRU / encoder-decoderAttention built directly on thisChapter 2. Words become vectors
  38. 2013word2vecELMo was influenced by thisChapter 2. Words become vectors
  39. 2011NLP from scratchELMo was influenced by thisChapter 1. Language as probability
  40. 2003Neural LMNLP from scratch was influenced by this
    word2vec was influenced by this    Chapter 1. Language as probability
  41. 1997LSTM Turning pointELMo built directly on thisChapter 1. Language as probability
  42. 1990Recurrent networksLSTM was influenced by thisChapter 1. Language as probability

After: what it led to

No documented descendants are recorded for Lost in the middle without weak links.

No documented descendants are recorded for Lost in the middle without weak links.

Lost in the middle has 8 recorded descendants, including indirect ones, across 3 chapters, reaching forward to Gemini 3.1 in 2026. Showing its direct descendants.

Lost in the middle has 8 recorded descendants, including indirect ones, across 3 chapters, reaching forward to Gemini 3.1 in 2026. Showing all of them.

  1. 2023Lost in the middleThis milestone.
  2. 2024Gemini 1.5was influenced by Lost in the middle (indirect influence)Chapter 10. Reading more than you remember
  3. 2024Gemini 2.0built directly on Gemini 1.5Chapter 14. Where the threads meet now
  4. 2025Gemini 2.5was influenced by Gemini 1.5Chapter 12. Thinking longer
  5. 2025IMO goldbuilt directly on Gemini 2.5Chapter 12. Thinking longer
  6. 2025IMO-ProofBenchwas influenced by IMO goldChapter 12. Thinking longer
  7. 2025Gemini 3was influenced by Gemini 1.5Chapter 14. Where the threads meet now
  8. 2026Gemini 3 Deep Thinkbuilt directly on Gemini 3Chapter 12. Thinking longer
  9. 2026Gemini 3.1built directly on Gemini 3Chapter 14. Where the threads meet now

Direct links and their sources

Where it came from

Influenced by

What it led to

Influenced